Weekly field note
Measuring AI Visibility When the Citation Set Keeps Moving
Published November 16, 2026
SISTRIX logged 56 percent weekly churn in AI Mode and 74 percent in ChatGPT, while AI Overviews held steady for 53 percent of prompts, so the figure worth reporting is a rate across runs

*Image: Example Control Chart - DSE Si Etch by DeminJanu, licensed under CC BY-SA 4.0.*
SISTRIX tracked 82,619 prompts and 1,548,213 weekly snapshots across six countries from 17 December 2025 to 8 April 2026, and its citation drift study reports that Google's AI Mode replaced 56 percent of its cited domains every week while ChatGPT Search replaced 74 percent. One week of answers is a single draw from a set that changes before the next report. The figure that survives comparison is a rate, meaning how often a domain appears across a fixed prompt set, on one surface, over several runs.
Profound compared 11 to 13 June 2025 with 11 to 13 July 2025 at roughly 80,000 prompts per platform, and reported monthly drift of 59.3 percent for Google AI Overviews, 54.1 percent for ChatGPT, 53.4 percent for Copilot and 40.5 percent for Perplexity, with a January-to-July range of 70 to 90 percent. Both studies are vendor measurements from proprietary tracking, so each percentage describes the platform and window its author chose.
The two disagree about which surface moves most. SISTRIX calls ChatGPT the most volatile and AI Overviews the most stable, while Profound measured slightly more monthly drift in AI Overviews. The definitions differ, weekly churn-in against a monthly set difference, so a drift figure is comparable only to another figure produced the same way.
What the drift numbers count
SISTRIX defines churn-in as the share of domains in a response that were absent the previous week and retention as the share carried over. The pair does not add to 100 because the number of domains per response moves. All of its figures compare domains, and the same post reports that URL-level drift runs higher, 85 percent against 74 percent in the comparison it gives. A site can hold a place in the citation set while the page cited changes underneath.
Profound measures a set difference across two monthly windows rather than a weekly chain. It compares prompts that appear in both windows, and SISTRIX states a narrower limit for its ChatGPT data: only prompts with consistent source attribution were included. Neither study is a controlled experiment on why a source dropped.
Why the three surfaces behave differently
Google describes the mechanism in its documentation on AI features: both AI Overviews and AI Mode may use query fan-out, issuing several related searches, and "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary." AI Overviews also appear only when Google's systems judge them additive to classic Search, so part of a run can return no Overview at all.
The citation sets barely overlap. SISTRIX found that for the same prompt, AI Overviews and AI Mode cite different domains 83 percent of the time, with a Jaccard overlap of 0.17, and that AI Mode and ChatGPT overlap at 0.125. Its explanation includes the language of the sources: in the German market, 80 percent of AI Mode core domains were German-language, while 68 percent of ChatGPT's core sources were English even for German queries.
The internal structure differs too. In AI Mode, 86.5 percent of prompts had a stable core of one to five domains, and the remaining domains rotated at 89 percent a week. AI Overviews showed no source change at all for 53 percent of prompts across the 17 weeks, while 19 percent rotated at 46 percent churn-in. For ChatGPT, the median prompt had no domain present in all 17 weeks, and each response cited 3 to 4 domains against 14 to 16 in AI Mode.
A platform-wide average hides that spread, and an average across platforms hides it further.
What tends to stay, and what to count
In Google AI Mode, brand queries were more stable but not fixed. The brand's own domain appeared for 43 percent of brand queries across all 17 weeks, and the 12 to 15 domains cited alongside it changed at 70 percent a week. Brand queries drifted 45 percent against that platform's 56 percent average, close enough that a brand should measure rather than assume.
Domain type predicted survival in AI Mode. SISTRIX's core rates were 24 percent for YouTube, 16 percent for large technology platforms, 12 percent for Wikipedia, 10 percent for marketplaces, 3 percent for forums and user-generated content, and 1.4 percent for news and media. The same hierarchy did not appear in ChatGPT, where no category held a clear advantage.
SISTRIX draws a practical line from this. Its domain-level numbers are the conservative ones, since drift at the URL level runs 15 percent higher. It treats domain presence as the measurable unit and describes the ranking of a single URL in AI responses as not meaningful. Count the domain, and treat "we are cited on this page" as a claim that needs a date.
Making two runs comparable
- Fix the prompt set before the first run and keep the wording. A changed prompt measures a different question.
- Keep one surface per series. SISTRIX found that AI Overviews, AI Mode and ChatGPT cite domains that differ by more than 80 percent for the same question, so a combined score mostly records their disagreement.
- Hold country and language constant. SISTRIX measured 54 to 59 percent weekly churn in AI Mode across six countries and wider variation in ChatGPT: 42 percent in France, 60 percent in the UK and 74 percent in Germany.
- Record the date and the product name. Google's documentation says the two surfaces may use different models and techniques, and a dated log keeps each run tied to the surface it measured.
- Count domains, not URLs, for the reason SISTRIX gives about single-page placement.
- Log crawl-policy changes. OpenAI's crawler documentation says a robots.txt update can take about 24 hours to reach its systems, and a site opted out of OAI-SearchBot will not be shown in ChatGPT search answers. A run on the day of a change measures the old access state.
- Report a rate over several runs. Profound argues that 40 to 60 percent drift makes single data points meaningless, and SISTRIX's equivalent advice is to base success metrics on presence over time.
Where an audit score fits
GeoAnalyzer's about page describes a score across five categories, Entity Clarity, Direct Answers, Trust Signals, Competitive Positioning and Technical Accessibility, built from signals its authors say AI engines consider when choosing sources to cite. Its pricing page lists a free 0-100 score with a Bronze, Silver or Gold tier and the top hesitation identified, then a $19 one-time report that adds a five-part breakdown, AI query simulations, a competitor landscape and a 30-day roadmap. The score reads mostly site-level inputs. GeoAnalyzer's how-it-works page lists schema markup, heading hierarchy, FAQ and answer blocks, crawlability and JavaScript rendering among the signals behind the five categories, and it also names off-site components: Entity Clarity checks consistent name, address and phone details across the web and knowledge-graph presence on Wikipedia, Wikidata and Crunchbase, while Competitive Positioning compares the site against competitors being cited. Those inputs set the conditions a site runs under, while the citation rate is the outcome measured across runs.
The pricing page describes the simulations as live simulated queries showing whether AI would cite you, your estimated position, the response snippet and the competitors alongside you. A report is a point-in-time check, which makes those simulations useful as a demonstration and limited as a trend. On 16 November 2026 the same site's homepage sold a $25 AEO Mastery course and listed an Appearance Tracker described as logging runs to "get the rate that survives randomness", while the pricing page still listed the $19 report, so confirm which offer is live before paying.
Content changes can shift the underlying probability. The GEO benchmark paper, accepted to KDD 2024, reported that content-level optimization methods raised visibility by up to 40 percent in generative engine responses across its GEO-bench queries, with effects that varied by domain. A page-level edit and platform drift are separate sources of movement, and a single audit cannot separate them.
The number to put in front of a client
Take a 40-prompt set run weekly for eight weeks, which gives 320 prompt-runs. If your domain appears in 96 of them, the rate is 30 percent. Report the numerator and the denominator with the surface, prompt set, country and run dates, because a rate without them cannot be checked.
If a movement sits inside the drift band the studies measured, collect more runs before calling it a change, and do not attribute it to the last edit. Neither study separates a change in retrieval policy from sampling variation, and the two methods differ enough to rank the platforms differently. One check in SISTRIX argues against a purely mechanical artifact for stable AI Overviews prompts: in 87 percent of the AI Overviews prompts whose sources did not change, the generated response text changed week to week while the cited domains stayed identical.
Choose the surface and the prompt set you will judge, log every run, and compare rates across at least several weeks.